Deep Cogito has released Cogito V2, a new family of open source AI models that hone their own reasoning skills.
The new Cogito V2 lineup, released under the open source license, includes four hybrid inference AI models. Two medium sizes, 70B and 109B parameters, two large versions in 405B and 671B.
The largest 671B Mixed Mixture (MOE) model is already touted as one of the most powerful open source AIs in the world. The company reports that it is competing with DeepSeek’s latest news, filling gaps in its own systems, such as the O3 and Claude 4 Opus.
But the real story isn’t merely about size or power. It is about fundamental changes in AI learning methods. Cogito V2 is designed to internalize your own inference process, not only “think” over time to find answers.
This internalized inference is achieved by a technique called iterative distillation and amplification (IDA) that distills discoveries from search to the core parameters of the model. The goal is to build a stronger “intuition” so that the model can predict the outcome of its own inference without performing the entire search.
The open source AI model has a better “gut feeling” to a good approach, so its inference chain is 60% shorter than its rival chains like DeepSeek R1.
This efficiency extends to budget. Deep Cogito says it has developed all models, totaling under $3.5 million, from experiments to final training. There are still a lot of money for you or me, but only a small amount compared to many major AI lab spending.
The flagship 671B model attracted special attention and was trained to not only improve the final answer, but also to improve the thought process itself. This approach discourages the model from being “wrenched” and rewards a more direct path to the solution. Performance data suggests that Deep Cogito’s open source AI model matches or exceeds the latest DeepSeek version on key benchmarks.
Perhaps one of the most surprising results is the ability of models to infer about images. A skill they never explicitly trained.
The team shared an example of this reasoning in which a deep cogget open source AI model compares two images of ducks and lions, showing a deep thought process about habitat, color and composition purely through transfer learning. Deep Cogito believes this emergency property could become a powerful way to bootstrap training data for future multimodal inference systems.
Looking ahead, the Deep Cogito team plans to “hill climb the benefits of repetitive self-improvement” in their quest to build super intelligence. They revised their commitment to making every AI model they create open source.
reference: The leak suggests that Openai’s open source AI model release is imminent
Want to learn more about AI and big data from industry leaders? Check out the AI & Big Data Expo in Amsterdam, California and London. The comprehensive event will be held in collaboration with other major events, including the Intelligent Automation Conference, Blockx, Digital Transformation Week, and Cyber Security & Cloud Expo.
Check out other upcoming Enterprise Technology events and webinars with TechForge here.
